基于贝叶斯模型平均和遗传粒子滤波的积雪数据同化方法研究

The proposed research project will study how to improve the accuracy of snow depth prediction on basis of multi-model (CoLM, SiB2, Noah) and multi-observation (snow depth in-situ, ground-based microwave radiometer, MODIS snow cover area, and AMSR brightness temperature) obtained in Northern Xinjing. A novel snow data assimilation framework (named BMA-GPF) will be developed to realize the integration of multi-model and multi-observation, which is based on Bayesian Model Averaging (BMA) and Genetic Particle Filter (GPF). The BMA is used for multi-model ensemble prediction of snow depth, which has the capacity to reduce the uncertainties of model structure. In order to improve the accuracy of snow depth prediction, the GPF is also adopted in this project to realize the joint assimilation of multi-observation (snow cover area and brightness temperature). Additionally, the problem of particle degeneracy in particle filter is also discussed in this project. The expected research results can provide a new idea for snow data assimilation, promote the application of multi-source remote sensing data in the field of hydrology, enhance the capability of snow depth forecast.

本项目以北疆地区多源数据（站点雪深、地基微波辐射计、MODIS积雪面积、AMSR微波亮温）为基础，实现多陆面过程模型（CoLM、SiB2、Noah）对雪深的模拟。在此基础上，发展基于贝叶斯模型平均和遗传粒子滤波（BMA-GPF）的积雪数据同化框架，实现多模型和多观测的综合集成，从模型、数据、算法三方面探索提高雪深模拟精度的途径。将利用贝叶斯模型平均方法实现雪深的多模型集合预报，减少积雪模型的不确定性；利用遗传粒子滤波算法实现多源遥感观测（积雪面积和微波亮温）的联合同化，解决粒子退化问题，提高雪深预报精度。本项目的开展为积雪数据同化研究提供了新的思路，将促进多源遥感数据在水文领域的应用，提升雪深的模拟和预报水平。

本项目以我国典型的季节性积雪区域—北疆为研究区，重点围绕Noah-MP模型，开展了多模型模型参数化方案的敏感性评估，在此基础上构建了结合GPF（遗传粒子滤波）和BMA（贝叶斯模型平均）算法的积雪数据同化框架。在积雪面积产品重建、积雪面积数据同化、积雪模型多参数化方案评价方面取得了重要成果。具体包括：（1）评价了Noah-MP模型积雪参数方案的敏感性与不确定性；（2）构建了基于机器学习的“积雪面积-雪深”关系曲线，并利用北疆观测数据进行了验证；（3）构建了结合GPF（遗传粒子滤波）和BMA（贝叶斯模型平均）算法的积雪数据同化框架；（4）发展了基于机器学习的积雪面积比例产品重建、积雪产品去云、以及雪深反演方法。在项目执行期间，共发表论文10篇（SCI论文9篇，中文核心论文1篇），培养博士2名。